A rate limit caps how many requests or actions a system permits in a specified period. In AI systems, it provides a containment layer against abuse, accidental loops, and runaway agent behavior. section 4.7.1 section 4.9 section 2.4.1
Set action caps per minute, session, or day according to the risk of the operation. A limit on frequency does not bound the impact of each action, so pair it with scope limits on records affected, transaction value, or distinct resources modified. section 4.7.1
Refund containment
If an AI assistant can process only 10 refunds per hour, a misinterpretation can trigger at most 10 incorrect refunds in that hour. A separate transaction-value cap limits the amount affected. section 4.7.1
Containing a request loop
A QA engineer accidentally sent thousands of requests in a loop, incurring several hundred dollars before anyone noticed. The team added per-session rate limits, token caps, and hard spend limits connected to a finance dashboard. section 4.9
Rate limits complement, rather than replace, other safeguards. Use approval and explicit exclusions for consequential actions, and log the intent chain for state-changing operations so investigators can distinguish a user request from the AI’s interpretation. section 4.7.1


















